Computer Vision Beyond Surveillance: Industrial Perception for Manufacturing QC
How Araskova's Vigil system applies semantic segmentation and anomaly detection to factory production lines — moving beyond CCTV surveillance into real-time quality intelligence.
The Problem With Industrial QC Today
Most factory quality control is still human-eye inspection, end-of-line sampling, or rigid rule-based machine vision (blob detection, template matching). These approaches share a common failure mode: they catch defects after production, not during it.
By the time a batch reaches inspection, hundreds of defective units may already exist. The cost is not just the defective units — it is the root-cause latency. You cannot trace a defect back to the exact machine state, temperature, or material batch that caused it, because you weren't watching continuously.
Vigil's Architecture
Vigil is Araskova's edge-compute perception node for factory floors. It sits on the production line, not at the end of it.
Core Components
1. Streaming Ingestion
Industrial cameras (GigE Vision or USB3 Vision) feed directly into the edge node at 60–120fps. No frames are dropped; the system operates on ring buffers with configurable retention windows.
2. Inference Layer (Powered by RUSTAMA)
We run lightweight semantic segmentation models — trained specifically on the client's production line defect taxonomy — using our RUSTAMA Rust-native inference runtime. RUSTAMA was built for exactly this: deterministic latency (<20ms per frame), no JVM or Python interpreter overhead, and optional cloud connectivity for complex reasoning.
3. Anomaly Envelope
Rather than binary pass/fail, Vigil builds a statistical envelope of "normal" surface texture, geometry, and color for each product variant. Deviations are scored and timestamped against production parameters (conveyor speed, line temperature, batch ID).
4. Root-Cause Correlation
The anomaly log is joined with the production data stream (via OPC-UA or Modbus integration, depending on the PLC). This means when a cluster of defects appears, the operator can immediately see: belt speed was 3% high, temperature was 2°C above setpoint, and this batch of material was involved.
Why Edge — Not Cloud?
Three reasons:
- Latency: A 100ms cloud round-trip is invisible in a dashboard. It is catastrophic on a production line running at 120fps where a defect needs to trigger a reject gate in <50ms.
- Data sovereignty: Factory floor imagery contains proprietary product geometry, surface finishes, and tooling marks. This is trade-secret data that cannot leave the facility.
- Connectivity: Many manufacturing facilities in Kerala (and globally) have unreliable or bandwidth-constrained internet. RUSTAMA's on-device inference means Vigil works with zero connectivity.
Status
Vigil is in active development. If you operate a manufacturing facility and want to run a pilot, contact us.